💻 computer science

Beyond Implicit Force: Evaluating Explicit Force-Torque Proxies in Action Chunking with Transformers

This paper demonstrates that while Action Chunking with Transformers (ACT) often relies on implicit teleoperation tracking errors to achieve contact-rich manipulation, replacing these hidden cues with explicit joint-torque proxies derived from onboard sensors effectively recovers or even enhances force-aware behavior across diverse real-world tasks.

Original authors: King Hang Wong, Lingqiao Liu, Feras Dayoub

Published 2026-07-17

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: King Hang Wong, Lingqiao Liu, Feras Dayoub

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to do delicate chores, like wiping a table or plugging in a charger. You can't just show the robot a video; it needs to feel what it's doing. But robots don't have skin, and cameras can't see "pressure" or "friction." This is the tricky world of contact-rich manipulation, where a robot has to interact with the physical world without breaking things or getting stuck. To teach robots, scientists often use a method called Imitation Learning, where the robot watches a human do a task and tries to copy them. A popular tool for this is called ACT (Action Chunking with Transformers), which is like a super-smart predictor that guesses the next few moves a robot should make all at once, rather than one step at a time. The big question everyone is asking is: How does the robot know when to push harder or back off if it can't actually "feel" the object? Does it need expensive, specialized sensors, or can it figure it out just by watching?

This paper dives into a surprising secret hidden in how we teach robots. The researchers found that the popular ACT robot was actually "cheating" a little bit. When humans teach these robots using a special "leader-follower" setup (where a human moves a master arm, and the robot copies it with a slave arm), the robot wasn't just copying the position of the arm. It was secretly learning from the struggle. If the robot's arm hit a wall and couldn't move as fast as the human's arm, that tiny lag or "tracking error" acted like a hidden signal, telling the robot, "Hey, something is in the way, push harder!" The paper asks: What if we take away that hidden struggle signal? Does the robot forget how to handle contact?

To find out, the authors built a version of the robot that ignores the human's "struggle" and only predicts where the robot's own arm should go. They called this ACT-o. When they tested it on real-world tasks like wiping a board, plugging in a charger, or pressing a soft bottle, the robot completely fell apart. Without that hidden "lag" signal, the robot became hesitant, froze, or just hovered above the object, failing to apply the necessary force. It was like a musician who could play the notes but forgot how to press the keys hard enough to make a sound.

However, the story doesn't end in failure. The researchers then tried a simple fix: they gave the robot a new sense using data it already had. Instead of expensive external sensors, they used the robot's own internal motor current (how much electricity the motors are using) as a "torque proxy"—a rough estimate of how much force the robot is exerting. When they fed this simple "force feeling" back into the robot, it came back to life. The robot could suddenly distinguish between a hard surface and a soft one, stop exactly when it touched a block of foam, and even perform gentle "taps" to align a plug.

The paper suggests that while the hidden struggle from the old teaching method was a powerful, accidental teacher, we don't need to rely on it anymore. By simply adding a basic estimate of force (derived from motor current) to the robot's brain, we can make it just as good, or even better, at handling contact than before. This means we can teach robots to do delicate, force-sensitive jobs even on cheap hardware that doesn't have fancy, expensive force sensors, as long as we give them a way to "feel" their own motors working.

Technical Summary: Beyond Implicit Force: Evaluating Explicit Force-Torque Proxies in Action Chunking with Transformers

1. Problem Statement

Robust contact-rich manipulation remains a significant challenge in robotics because critical interaction signals—such as contact forces, compliance, and friction—are often weakly observable through vision and kinematics alone. While Action Chunking with Transformers (ACT) has demonstrated strong performance in fine-grained manipulation, its success in real-world deployments is often intertwined with leader–follower teleoperation (e.g., ALOHA-style systems).

In these systems, the low-level controller tracks leader commands via position control. The resulting tracking error (the discrepancy between the commanded leader motion and the executed follower motion) implicitly encodes contact onset, resistance, and constraint violations. This paper investigates a critical dependency: To what extent does ACT rely on this hidden, teleoperation-induced interaction cue? Furthermore, it addresses the practical limitation that many platforms lack dedicated six-axis force/torque (F/T) sensors due to cost and integration complexity, raising the question of whether onboard torque signals (derived from motor currents or joint effort) can serve as effective explicit proxies to replace or enhance these implicit cues.

2. Methodology

The authors conducted a controlled real-world study on an ALOHA Solo platform (featuring kinematically identical WidowX-250s and ViperX-300s arms) to isolate and evaluate interaction signals.

A. Observation-Centric ACT (ACT-o)

To remove the implicit teleoperation cue, the authors introduced an observation-centric ACT variant (ACT-o).

Standard ACT: Predicts future leader joint positions ( $A_{leader}$ ). The model implicitly learns force dynamics through the discrepancy $\delta = q_{leader} - q_{follower}$ .
ACT-o: Predicts future follower joint states ( $Q_{follower}$ ) directly. This reformulation decouples the policy from the leader-follower lag, effectively removing the latent force-awareness signal while preserving the rest of the learning pipeline.

B. Explicit Torque Proxies

To recover force-awareness without external F/T sensors, the authors integrated explicit torque proxies derived from onboard motor current measurements ( $\tau = k_t I$ ).

Injection Method: Torque values are treated as "first-class" inputs. They are concatenated with proprioceptive joint states prior to the hidden-dimension projection, allowing the Transformer encoder to learn a mapping between torque transients and physical resistance.
Model Variants: Four configurations were evaluated:
1. Base ACT: Standard leader-target prediction (Implicit force via discrepancy).
2. ACT-o: Follower-target prediction (No implicit force).
3. ACT + $\tau$ : Base ACT with added torque inputs.
4. ACT-o + $\tau$ : Observation-centric prediction with added torque inputs.

C. Experimental Tasks

The evaluation spanned four contact-sensitive tasks designed to test different aspects of force interaction:

Board Wipe: Surface following and maintaining consistent normal force during lateral movement.
Plug Insert: High-precision, multi-mode alignment and insertion requiring force control.
Soft Bottle Press: Stiffness discrimination between visually identical but mechanically distinct states (high vs. low stiffness).
Foam Stop: Force-equilibrium behavior, stopping upon light resistance against a textureless object.

3. Key Results

The experiments revealed that removing the implicit teleoperation cue causes catastrophic failure in contact-rich tasks, while explicit torque proxies successfully restore and enhance performance.

Collapse of ACT-o: Removing the implicit cue (predicting follower states without torque) led to severe performance degradation. In the Board Wipe task, success dropped to 0% for ACT-o (vs. 60% for Base ACT) due to "hovering" behavior (failure to maintain downward force). In Plug Insert, ACT-o failed to initiate task dynamics or differentiate states, resulting in 0% success in some conditions.
Restoration via Torque Proxies: Adding torque proxies to the observation-centric model (ACT-o + $\tau$ ) recovered robust contact behavior.
- Board Wipe: Success rates improved from 15% (ACT-o) to 95% (ACT-o + $\tau$ ).
- Soft Bottle Press: The torque-augmented model successfully distinguished between high and low stiffness states, whereas ACT-o failed to differentiate them.
- Plug Insert: Torque-augmented policies exhibited "adjustment-taps" and gentle reactive motions, achieving 90% success compared to 30% for Base ACT and 0% for ACT-o.
Force as a Geometric Anchor: In tasks like Foam Stop, torque proxies allowed the policy to converge to a stable equilibrium force profile closely matching human demonstrations, even when visual cues were ambiguous.

4. Key Contributions

Isolation of Implicit Cues: The paper formalizes and ablates the teleoperation-induced discrepancy signal by introducing an observation-centric ACT target. It demonstrates that removing this cue collapses performance on contact-rich tasks, proving that ACT's success is partially dependent on this hidden artifact.
Explicit Torque Proxy Evaluation: The study provides systematic evidence that simple, onboard torque proxies (derived from motor current) can recover contact-aware behavior in the absence of external F/T sensors.
Task-Level Analysis: The authors provide quantitative results and diagnostic analyses characterizing reactivity and stability across diverse contact regimes (impact, friction, stiffness discrimination), offering practical guidelines for integrating torque proxies into ACT-style policies.

5. Significance and Claims

The paper claims that on real hardware, the implicit teleoperation cue is a recoverable source of force-awareness. Crucially, it demonstrates that where torque signals are available, a simple proxy matches, surpasses, or further enhances the performance of policies relying on implicit cues.

The authors argue that explicit force modeling is not merely beneficial for general policy modeling but is essential for robust force-aware tasks. By treating screw-axis torque as a first-class input, the proposed method enables robust manipulation on platforms lacking specialized hardware, bridging the gap between high-fidelity teleoperation data and practical, sensor-constrained deployment.

Limitations and Future Directions:
The authors note that joint-side torque estimates lack the high-frequency resolution of dedicated six-axis F/T sensors or optical tactile arrays. Future work is proposed to explore cross-platform proxy transfer, force-aware distillation (combining VLMs with force-integrated ACT), and active sensing strategies where policies explicitly probe objects to disambiguate visual uncertainty.